Bus line network intelligent scheduling method based on machine learning model

Abstract

The application discloses a bus line network intelligent scheduling method based on a machine learning model, belongs to the technical field of data processing, and comprises the following steps: receiving line network sensing data; when it is detected that the line network sensing data has an abnormal evolution trend, generating a line network operation situation snapshot and conveying the line network operation situation snapshot to a scheduling information coordination processing unit to generate each scheduling information entry; calling each scheduling information entry to perform overlap detection, outputting an overlap relationship matrix, and dividing to obtain a normal scheduling information set and each conflict scheduling information set; obtaining machine learning model training parameters of each conflict scheduling information entry, analyzing to obtain a model history stability index and a model generalization confidence index of the machine learning model, and screening each conflict scheduling execution instruction according to the model history stability index and the model generalization confidence index to execute the bus line network intelligent scheduling, so that the technical problem that different scheduling information has deviation and inconsistency in the state cognition level in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for intelligent scheduling of public transport networks based on machine learning models. Background Technology

[0002] With the continuous expansion of urban public transportation networks and the dynamic changes in travel demand structure, the operation status of bus networks exhibits obvious time-varying and coupling characteristics. Therefore, it is necessary to realize intelligent scheduling of bus networks based on multi-dimensional operation data.

[0003] Existing technologies typically optimize bus departure intervals, operating rhythms, and regional capacity allocation by constructing data-driven scheduling models. For example, by collecting passenger flow data, station information data, arrival and departure data, and vehicle departure data, waiting time models and interpolation models are built to characterize passenger waiting times and inter-station speeds at different times. Then, based on multi-agent reinforcement learning algorithms, passenger waiting times, station segment speeds, and planned departure intervals are jointly trained to obtain the departure times of each vehicle on each route, thus optimizing bus scheduling. Alternatively, by judging the state of the time period, when peak hours are identified, the speed of buses in each area is obtained, and congestion coefficients are constructed to generate multiple scheduling schemes. These schemes are then simulated and tested using a digital twin model of vehicle scheduling. An improvement index is constructed based on changes in the congestion coefficient, allowing the most feasible optimal scheduling scheme to be selected from multiple options.

[0004] For example, Chinese invention patent CN116307448B discloses a method for intelligent bus scheduling based on multi-agent reinforcement learning, which includes: constructing a waiting time model and an interpolation model based on passenger flow data, station information data, arrival and departure data, and vehicle departure data; obtaining the waiting time of each passenger based on the waiting time model, and obtaining the speed between stations at different time periods based on the interpolation model; training the waiting time of each passenger and the speed between stations at different time periods and the planned departure interval based on the multi-agent reinforcement learning algorithm to obtain the final departure interval, and thus obtain the departure time of each vehicle on each route, thereby realizing bus vehicle scheduling.

[0005] The above-mentioned technology has at least the following technical problems:

[0006] Existing technologies generally model and analyze multi-source data such as passenger flow characteristics, operating speed, and regional status, and use reinforcement learning models, regional status models, or digital twin models to generate scheduling decisions to achieve intelligent scheduling of public transport networks. However, in the above technical solutions, the generation process of scheduling information and the coordination and verification process between scheduling information usually present a serial coupling processing mode in the system architecture. When multiple scheduling information is generated in a short period of time, due to factors such as inconsistent timestamps of multi-source data collection, delays in model inference and decision calculation, and lag in feedback of scheduling execution results, the scheduling information generated and executed earlier has objectively changed the actual operating status of the public transport network. Subsequent scheduling information is still generated and judged based on the status data of the previous moment or historical prediction results, resulting in deviations and inconsistencies in the state cognition of different scheduling information. Summary of the Invention

[0007] To address the technical problem in existing technologies where the asynchronous updating of network operation status and the generation of scheduling decisions lead to discrepancies and inconsistencies in the state perception of different scheduling information, this invention provides a smart scheduling method for public transport networks based on a machine learning model. The technical solution is as follows:

[0008] A machine learning-based intelligent bus network scheduling method includes: receiving network sensing data; when an abnormal evolution trend is detected in the network sensing data, generating a snapshot of the network operation status and sending it to the scheduling information coordination and processing unit to generate various scheduling information terms; retrieving each scheduling information term for overlap detection, outputting an overlap relationship matrix, and dividing it into a normal scheduling information set and a conflict scheduling information set, the conflict scheduling information set including conflict scheduling information terms with a certain overlap relationship; obtaining the machine learning model training parameters for each conflict scheduling information term, analyzing the model's historical stability index and model generalization reliability index, and selecting conflict scheduling execution instructions accordingly to execute intelligent bus network scheduling; the machine learning model training parameters include the model's historical stability parameter and model generalization reliability parameter; managing the executed conflict scheduling execution instructions with scheduling effectiveness identifiers, and controlling the inclusion and exclusion of scheduling information terms in subsequent scheduling cycles based on the scheduling effectiveness identifiers to complete intelligent bus network scheduling.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0010] 1. The intelligent scheduling method for bus networks based on machine learning models provided by this invention performs time alignment and state field evolution analysis on network perception data within a preset scheduling cycle, and generates a snapshot of the network operation status when abnormal evolution trends are detected. This generates various scheduling information entries on the basis of unified state cognition, thereby realizing the synchronous association between the scheduling decision generation process and the network operation status evolution process. This effectively solves the technical problem in the prior art where the asynchronous updating of network operation status and scheduling decision generation leads to deviations and inconsistencies in the state cognition level of different scheduling information.

[0011] 2. This invention analyzes the resource identification dimension of each scheduling information term to construct an overlapping relationship matrix, and divides the normal scheduling information set and conflict scheduling information set accordingly. This allows for the explicit identification of potential scheduling resource conflict relationships before scheduling execution, thereby achieving structured management and grouping of conflict scheduling information. This effectively avoids the problems of line resource competition, timing conflicts and uncontrollable scheduling effects caused by the superimposed execution of scheduling instructions in the prior art.

[0012] 3. This invention obtains the model's historical stability index and model generalization credibility index corresponding to each conflict scheduling information term, and combines a preset threshold and a comprehensive confidence value calculation mechanism to perform multi-dimensional credibility screening and sorting of candidate scheduling information terms. In this way, the scheduling execution instruction with the highest execution reliability is selected in the conflict scheduling scenario, thereby achieving joint constraints on the predictive stability and cross-cycle consistency of the machine learning model.

[0013] 4. This invention manages the scheduling effectiveness flags of executed conflict scheduling instructions and introduces scheduling coordination freeze constraints in subsequent scheduling cycles to control the merging and exclusion of scheduling information terms. This ensures that all scheduling information terms are generated based on the same network operation status snapshot, thereby achieving consistency and traceability of scheduling decisions across scheduling cycles. It effectively solves the problems of repeated correction of scheduling results, repeated triggering of overlapping detection, and frequent oscillation of scheduling strategies in the prior art. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A macroscopic flowchart of the intelligent scheduling method for public transport networks based on machine learning models provided in the embodiments of this application;

[0016] Figure 2A flowchart illustrating the overall process of the intelligent scheduling method for public transport networks based on a machine learning model, as provided in this application embodiment.

[0017] Figure 3 A schematic diagram of the evolution trend curve and abnormal segments of the network status field of the intelligent scheduling method for bus networks based on machine learning models provided in the embodiments of this application;

[0018] Figure 4 A flowchart illustrating the abnormal evolution trend detection and network operation status snapshot generation process of the intelligent bus network scheduling method based on machine learning models provided in this application embodiment;

[0019] Figure 5 Heatmap of overlapping relationship matrix of scheduling information entries for the intelligent scheduling method of public transport network based on machine learning model provided in the embodiments of this application;

[0020] Figure 6 A schematic diagram illustrating the stability and reliability distribution of each candidate scheduling information term in the intelligent scheduling method for public transport networks based on a machine learning model provided in this application embodiment;

[0021] Figure 7 The flowchart illustrates the conflict scheduling execution instruction filtering and determination process of the intelligent bus network scheduling method based on a machine learning model provided in this application embodiment. Detailed Implementation

[0022] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0023] In embodiments of the present invention, words such as "examplely" and "for example" are used to indicate that something is being used as an example, illustration, or explanation. Furthermore, in embodiments of the present invention, the meaning of "and / or" can be both, or either one.

[0024] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0025] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0026] First embodiment, such as Figure 1 As shown, Figure 1The following is a macroscopic flowchart of the intelligent bus network scheduling method based on a machine learning model provided in this application embodiment. The method includes the following steps: receiving network sensing data; when an abnormal evolution trend is detected in the network sensing data, generating a snapshot of the network operation status and sending it to the scheduling information coordination and processing unit to generate various scheduling information terms; retrieving various scheduling information terms for overlap detection, outputting an overlap relationship matrix, and dividing the data into a normal scheduling information set and a conflict scheduling information set, wherein the conflict scheduling information set includes various conflict scheduling information terms with a certain overlap relationship; obtaining the machine learning model training parameters of each conflict scheduling information term, analyzing the model historical stability index and model generalization reliability index of the machine learning model, and selecting various conflict scheduling execution instructions accordingly to execute intelligent bus network scheduling; the machine learning model training parameters include the model historical stability parameter and model generalization reliability parameter of the machine learning model; managing the scheduling effectiveness identifiers of the executed conflict scheduling execution instructions, and controlling the incorporation and exclusion of scheduling information terms in subsequent scheduling cycles based on the scheduling effectiveness identifiers to complete intelligent bus network scheduling.

[0027] Figure 2 The following is a flowchart of the overall process of the intelligent bus network scheduling method based on a machine learning model provided in this application embodiment. When the network sensing data is determined to have an abnormal evolution trend after time alignment and state field evolution analysis, the system immediately generates a snapshot of the network operation status for the corresponding scheduling period and generates various scheduling information terms based on the snapshot. Subsequently, overlap detection is performed on all scheduling information terms to construct an overlap relationship matrix and divide it into a normal scheduling information set and multiple conflict scheduling information sets. For each conflict scheduling information set, the corresponding machine learning model training parameters are obtained, and the historical stability index and generalization confidence index of the model are analyzed. Through threshold screening and comprehensive confidence value calculation, the execution instructions for each conflict scheduling are determined. After the scheduling execution is completed, the scheduling effective identification management is performed on the executed and unexecuted conflict scheduling execution instructions, and when entering the subsequent scheduling period, the merging and exclusion of scheduling information terms are controlled based on the scheduling coordination freeze constraint.

[0028] It should be noted that machine learning algorithm models are not limited to a single model structure, but rather a collection of models used to predict, evaluate, and analyze the credibility of bus network scheduling information terms. These models may include, but are not limited to: time series prediction models (such as regression structure sequence models, state space models, etc.) used to characterize the continuous changing trends and historical evolution of scheduling results; and supervised learning models (such as gradient boosting tree models, random forest models, support vector regression models) used to learn nonlinear mapping relationships from multidimensional scheduling features. The above models can be used individually or collaboratively to form a scheduling decision model system through ensemble learning.

[0029] Furthermore, the specific detection method for abnormal evolution trends is as follows: Within a preset scheduling cycle, the network sensing data after time alignment processing is analyzed for state field evolution. These state fields include departure interval, segment operation density, station load change, and capacity configuration status. The segment operation density field characterizes the degree of vehicle aggregation and operational saturation within a specific road segment per unit time. The station load change field characterizes the changing trend of station passenger flow load within the scheduling cycle. The capacity configuration status field characterizes the deviation between the line capacity allocation and the planned capacity within the current scheduling cycle. The magnitude and direction of change of each network state field between adjacent scheduling cycles are evaluated to obtain the network status data. The evolution results of the status fields include the change magnitude result and the change direction determination result. When the change magnitude result of at least one network status field exceeds the corresponding preset stable range and the change direction determination result is positive evolution or oscillating evolution (i.e., there is an abnormal change magnitude result of at least one network status field and the change direction determination result is positive evolution or oscillating evolution) and / or the network status fields of multiple lines show a coordinated abnormal offset in the same scheduling cycle and the change direction determination result is positive evolution or oscillating evolution, it is determined that there is an abnormal evolution trend in the network sensing data. The change direction determination result can be positive evolution (deterioration trend amplification), negative evolution (deterioration trend shrinkage, i.e., deterioration trend improvement) or oscillating evolution.

[0030] In this embodiment, it should be noted that the state field evolution analysis is not a single-point numerical judgment, but a time-series comparison analysis across scheduling cycles. Its core is to align and compare each state field in the current scheduling cycle with the corresponding field in the previous scheduling cycle, thereby depicting the changing trend of the network state.

[0031] The variation amplitude of each network status field between adjacent scheduling cycles (where variation amplitude characterizes the intensity of status field changes) is specifically calculated by performing absolute difference processing on the values ​​of the same status field within adjacent scheduling cycles to obtain the variation amplitude value of that field on the scheduling cycle scale. The direction of change characterizes the trend of state evolution, specifically including: positive evolution (a continuous increase or worsening trend in variation amplitude), negative evolution (a continuous decrease or mitigation trend in variation amplitude), or oscillatory evolution (unstable direction). It should be noted that the change direction determination is only performed when the variation amplitude result is abnormal; otherwise, it is not performed, meaning it is considered normal. The variation amplitude result specifically includes normal and abnormal variations. The variation amplitude of the network status field between adjacent scheduling cycles is compared with the corresponding variation amplitude threshold. If it is less than the threshold, the variation amplitude result is normal; otherwise, it is abnormal.

[0032] By performing state field evolution analysis on time-aligned network sensing data within a preset scheduling cycle, and evaluating the magnitude and direction of change of each network state field between adjacent scheduling cycles, an evolution result containing both magnitude and direction determination results is obtained. This achieves a quantitative characterization and directional discrimination of the dynamic changes in the operational status of the public transport network. Based on this, by comparing the evolution result with a preset stable state interval, and combining this with a comprehensive judgment on whether multiple routes exhibit coordinated abnormal offsets within the same scheduling cycle, the trend of network operational status evolving from stable to abnormal can be identified in advance before scheduling information is generated. This effectively avoids misjudgments caused by relying solely on instantaneous thresholds or single indicators. Consequently, the subsequently generated network operational status snapshot can accurately reflect the evolutionary background of the network status.

[0033] like Figure 3 As shown, Figure 3The diagram illustrates the evolution trend curves and abnormal segments of the network status fields in the intelligent bus network scheduling method based on a machine learning model provided in this application. The four network status fields (departure interval, section operation density, station load change, and capacity configuration status) can be understood as "taking a statistical snapshot of the network operation status at the end of each scheduling cycle," and then concatenating adjacent snapshots to form the evolution trend curve. The departure interval characterizes the time difference between two adjacent buses passing through the same observation point (station or section) on the same route, reflecting service uniformity and passenger waiting levels. A sustained increase or intensified fluctuation in this value often indicates delays, congestion, or uneven vehicle operation. Section operation density characterizes the degree of vehicle aggregation in key sections; an abnormal increase often corresponds to vehicles queuing or being blocked within the section, while an abnormal decrease corresponds to service disruptions or insufficient supply. Station load change characterizes the change in passenger backlog or congestion at a station within a scheduling cycle; a sustained increase usually indicates increased demand-side pressure or insufficient capacity supply. The capacity configuration status is used to characterize the level of effective capacity available on the supply side, such as the proportion of available vehicles. A decline in this status often indicates vehicle malfunctions, turnover obstructions, or resource shortages. The evolution during a cycle is shown by connecting adjacent period points with a broken line, where each point represents a snapshot statistical value for that cycle. The slope of the line represents the magnitude (absolute change) and direction (upward / downward) of the change from the previous cycle to the next, facilitating the detection of abnormal evolution trends. The horizontal axis in the graph represents the scheduling cycle number, and the actual duration of each scheduling cycle (e.g., 5 minutes) can be configured by system parameters. The vertical axis represents the field values ​​of the network status fields. The light-colored bands in the graph can be interpreted as a pre-defined "state stability interval" for each field under historical benchmarks or rule constraints, used to compare and determine whether the field values ​​in the current cycle have exceeded the limits or shown continuous deviation. When coupled changes (multi-field linkage) occur within a certain period or adjacent periods, for example, if "increased operating interval (slower service) + increased station load (passenger backlog) + increased section density (vehicles congregating in key sections) + decreased capacity allocation (insufficient available capacity)" is observed in the 7-8 period of scheduling, it indicates a coordinated abnormal deviation of supply and demand imbalance accompanied by operational obstruction; and if "station load and section operating density gradually decline" after scheduling period 9, and normal operation resumes in scheduling period 10 and beyond, it indicates that the network situation gradually returns to a stable range after the scheduling instructions take effect.

[0034] like Figure 4 As shown, Figure 4The flowchart for detecting abnormal evolution trends and generating network operation status snapshots in the intelligent bus network scheduling method based on machine learning models provided in this application embodiment is as follows: First, the multi-source network sensing data from different functional modules is timestamped and time aligned, and historical and forward data that do not belong to the current scheduling cycle are removed. Then, the state field evolution analysis is performed on the aligned network sensing data, and the change magnitude and direction of each network state field between adjacent scheduling cycles are calculated, and the corresponding evolution results are obtained. When the evolution result of at least one network state field exceeds its preset stable range, or when the network state fields of multiple lines show coordinated abnormal offsets in the same scheduling cycle, it is determined that there is an abnormal evolution trend in the network sensing data. The system performs state abstraction processing on the unified network state fields in the current scheduling cycle, constructs a basic dataset of network operation status, and solidifies this dataset into a network operation status snapshot for subsequent generation and analysis of scheduling information entries.

[0035] The network sensing data (departure interval data, section operation density data, station load change data, capacity configuration status data, and other vehicle data, as well as passenger data) are mapped to a unified scheduling time window by timestamps. The specific method is as follows: using a preset scheduling cycle as the time base, the start and end times corresponding to the current scheduling cycle are defined. This time window serves as the sole time base for network operation status analysis and snapshot generation within the current scheduling cycle. Using the unified scheduling time window, the network sensing data from different functional modules are uniformly parsed to generate timestamps. The timestamps reflect the moment when the network status represented by the corresponding sensing data occurs. The timestamps of each network sensing data are matched with the unified scheduling time window. When a data timestamp falls within the current scheduling time window, the data is mapped and aggregated into the data set of the current scheduling cycle. This completes the mapping of network sensing data to the unified scheduling time window by timestamps.

[0036] Historical data refers to data whose timestamp is earlier than the start time of the current scheduling cycle. Forward-looking data refers to data with future timestamps that some functional modules may send in advance due to differences in data collection, caching, or uploading mechanisms (such as pre-calculated operating parameters or time drift data caused by cache refresh). In other words, forward-looking data refers to data whose timestamp is later than the end time of the current scheduling cycle, and the network state it represents has not yet actually occurred within the current scheduling cycle.

[0037] The process of abstracting the state of time-aligned multi-source network sensing data and extracting unified network state fields includes the following steps: Classifying the raw network sensing data (unprocessed network sensing data, such as vehicle arrival / departure data, station load status data, and other vehicle and passenger data) output by different functional modules. For example, classifying vehicle arrival / departure data and departure time data into operational rhythm information. Mapping the classified data to preset network state dimensions, such as departure interval fields, segment operation density fields, station load change fields, and capacity configuration status fields. Unifying the processing of data from different data sources under the same state field, forming a corresponding state field value for each network state field within the current scheduling cycle.

[0038] A basic dataset of network operation status is constructed and then solidified into a snapshot of the network operation status for the corresponding scheduling period. Specifically, the following methods are employed: The network status fields extracted within the current scheduling period and their corresponding values ​​are organized to form a complete set of network operation status descriptions; each status field is encapsulated into a basic dataset of network operation status according to a pre-defined data structure, which uniformly describes the overall network operation status within the current scheduling period; the basic dataset of network operation status is bound to the corresponding scheduling period identifier and stored in a fixed manner, ensuring its immutability within the current scheduling period; the solidified snapshot of the network operation status serves as the sole state input for generating scheduling information terms, performing conflict detection, and scheduling coordination, preventing state baseline drift caused by real-time data changes within the same scheduling period. This complete processing yields an immutable scheduling snapshot, which is the snapshot of the network operation status.

[0039] By uniformly receiving network sensing data from different functional modules within a preset scheduling period and mapping it to the same scheduling time window based on timestamps, followed by time alignment and irrelevant data removal, the time deviation introduced by different collection frequencies and inconsistent transmission delays of multi-source sensing data can be eliminated in the time dimension. This ensures that the data entering the analysis stage strictly corresponds to the actual operating status of the same scheduling period. Furthermore, by performing state abstraction processing on the time-aligned multi-source network sensing data, the original, heterogeneous sensing data is uniformly transformed into standardized network status fields such as departure interval, section operating density, station load change, and capacity configuration status. This achieves a structured expression of the complex network operating status, avoiding the fragmented state understanding problem caused by subsequent scheduling decisions directly relying on scattered original data.

[0040] The network status field is constructed as the basic dataset of network operation status and solidified as a snapshot of network operation status for the corresponding scheduling period. This ensures that the network operation status within the scheduling period is completely recorded and referenced in a stable, reusable, and unchangeable form, providing a consistent status benchmark for the generation of subsequent scheduling information entries, conflict detection, and scheduling decision filtering.

[0041] Furthermore, the overlap relationship matrix is ​​output. The specific method is as follows: each scheduling information term is parsed according to the resource identification dimension, which includes the target dimension, route dimension, and time window dimension; based on the target dimension, a mapping table of the interaction relationships between scheduling information terms is constructed; when any two scheduling information terms have an intersection in at least one target dimension, they are marked as overlapping (where scheduling information terms refer to the scheduling information content that needs to be executed calculated by the existing model; the existence of an intersection means that a certain keyword appears repeatedly, such as the existence of the same vehicle or the same stations, etc.); the overlap relationships between all scheduling information terms are represented in matrix form to obtain the overlap relationship matrix.

[0042] In this embodiment, resource identifier dimension parsing is performed on each scheduling information term. Specifically, each scheduling information term undergoes semantic structure parsing, breaking it down into an object dimension, a route dimension, and a time window dimension. The object dimension is obtained by parsing the explicitly pointed scheduling entity in the scheduling information term, representing the routes, stations, vehicles, or capacity units directly affected by the scheduling behavior. The route dimension is obtained by reading the associated route code in the scheduling information term, limiting the spatial scope of the scheduling behavior. The time window dimension is obtained by parsing the corresponding effective time period in the scheduling information term, describing the range of influence of the scheduling behavior on the time axis. The effective time range or applicable time period of the scheduling is extracted from the scheduling information term to form the time window dimension, describing the range of influence of the scheduling information within the scheduling cycle. A unified identifier encoding process is applied to each dimension, ensuring that different scheduling information terms have comparable identifier forms across dimensions, thus completing the resource identifier dimension parsing.

[0043] The specific steps for constructing a mapping table of interaction relationships between scheduling information terms are as follows: Using all scheduling information terms generated within the current scheduling cycle as the object, pairwise combinations are formed to create candidate comparison pairs. For each candidate comparison pair, a matching judgment is performed on the interaction object dimension to identify whether two scheduling information terms point to the same type or the same specific resource object (bus). When two scheduling information terms intersect in the interaction object dimension, the intersection relationship is recorded, and the intersection content is labeled. For each pair of scheduling information terms with interaction object intersection, an interaction relationship mapping entry is generated. The mapping entry includes at least the term pair identifier (i.e., the two scheduling information terms with an intersection relationship), the intersection dimension type, and other corresponding resource identifiers. All mapping entries are summarized to form a mapping table of interaction relationships between scheduling information terms, which centrally reflects the potential resource associations between scheduling information, thus obtaining the interaction relationship mapping table.

[0044] The specific processing steps to obtain the overlap relationship matrix are as follows: Using the total number of scheduling information terms within the current scheduling cycle as the matrix's row and column dimensions, each row and column of the matrix maps to a scheduling information term; each mapping relationship in the interaction relationship mapping table is converted into a relationship identifier (such as a term pair identifier) ​​at the corresponding row and column position in the matrix; when any two scheduling information terms intersect in at least one interaction object dimension, the corresponding position in the overlap relationship matrix is ​​marked as overlapping; otherwise, it is marked as non-overlapping; the generated overlap relationship matrix undergoes symmetry and integrity checks to ensure that the matrix structure is consistent with the set of scheduling information terms; the final generated matrix is ​​the overlap relationship matrix.

[0045] Furthermore, the normal scheduling information set and each conflict scheduling information set are obtained. Specifically, based on the overlap relationship matrix, the scheduling information terms are analyzed for correlation. Scheduling information terms that do not overlap with other scheduling information terms are assigned to the normal scheduling information set. Scheduling information terms that have at least one overlap relationship with each other (such as overlap in the object dimension, overlap in the line dimension, etc.) are merged into the same conflict scheduling information set, thus obtaining each conflict scheduling information set. The scheduling information terms in each conflict scheduling information set have a common conflict object.

[0046] In this embodiment, as Figure 5 As shown, Figure 5The heatmap of the overlapping relationship matrix of scheduling information terms in the intelligent scheduling method for bus network based on machine learning model provided in the embodiments of this application is shown. The horizontal axis of the graph represents term number j (T1~T10), and the vertical axis represents term number i (T1~T10). The coordinates themselves are discrete numbers. The matrix cell M(i,j) is a binary determination result. Only 0 and 1 appear in the graph: 1 indicates that term Ti and term Tj have an intersection (i.e., there is an overlapping relationship) in at least one resource identification dimension, and 0 indicates no intersection and no overlap.

[0047] By performing correlation analysis on scheduling information terms based on the overlap relationship matrix, scheduling information terms that do not overlap with other scheduling information terms are classified into the normal scheduling information set, and scheduling information terms that have at least one overlap with each other are grouped into the corresponding conflict scheduling information set. This enables the system to clearly distinguish between scheduling information terms that can be executed in parallel and those that require coordination decisions within the same scheduling cycle. This avoids executing logically mutually influential scheduling information simultaneously when timestamps are inconsistent or there are differences in state perception, which would lead to the network operation state being superimposed or covered by each other multiple times.

[0048] Furthermore, the historical stability indices of the machine learning models are obtained. Specifically, the historical stability parameters of the machine learning models corresponding to each scheduling information term within the historical scheduling period are acquired. These parameters include output distribution entropy (LS), maximum a posteriori probability (LP), local prediction gradient norm (LF), and latent space average distance (LY). A pre-defined historical stability benchmark set is obtained and compared with the historical stability parameters to obtain the historical stability processing values ​​for each model. These values ​​include the output distribution entropy processing value (τLS / LS), the maximum a posteriori probability processing value (LP / τLP), and the local prediction gradient norm processing value (τL). The historical stability benchmarks (τLY / LY) and the latent space average distance are multiplied by corresponding weights based on the historical stability values ​​of each model, and then summed to obtain the historical stability index used to characterize the predictive stability of the machine learning model. The historical stability benchmark set includes the output distribution entropy benchmark (τLS), the maximum a posteriori probability benchmark (τLP), the local prediction gradient norm benchmark (τLF), and the latent space average distance benchmark (τLY). The historical stability benchmark set is the minimum standard to ensure model stability. If any parameter (the historical stability parameter) is lower than the corresponding benchmark value, it indicates poor stability, which may lead to bias in the prediction results. The historical stability weights include the output distribution entropy weight, the maximum a posteriori probability weight, the local prediction gradient norm weight, and the latent space average distance weight.

[0049] In this embodiment, the output distribution entropy is obtained by calculating the predicted probability vector output by the model. For example, within a historical scheduling period, statistical analysis is performed on the predicted probability vectors output by the machine learning model corresponding to each scheduling information term. The predicted probability vector is the set of normalized probability values ​​output by the model for each candidate scheduling action or category. Based on the predicted probability vector, logarithmic operations and weighted summation are performed on each probability value according to the information entropy calculation formula to obtain the output distribution entropy parameter used to characterize the uncertainty of the model output. The maximum a posteriori probability can be directly obtained from the model output layer. It is obtained by selecting the probability value corresponding to the prediction result with the highest probability from the posterior probability distribution of the model's output when predicting scheduling information terms. The input data of this model is wire mesh sensing data, used to reflect the model's confidence in a certain prediction result under given input conditions. The higher the maximum a posteriori probability, the higher the model's internal confidence in the current prediction conclusion. The local prediction gradient norm is used to characterize the sensitivity of the model output to small perturbations in the input. A larger local prediction gradient norm indicates a more sensitive model prediction to input changes and weaker prediction stability. The local prediction gradient norm can be calculated through the model's backpropagation mechanism. Specifically, within historical scheduling cycles, the local gradient of the input features (feature vectors of the network state field) corresponding to scheduling information terms is calculated at the model output, and the norm of the gradient vector is calculated. The latent space average distance measures the consistency of the model's representation of similar scheduling scenarios in the latent space. A larger latent space average distance indicates a more dispersed internal understanding of similar inputs and weaker stability. The latent space average distance can be obtained through the model's intermediate layer feature output interface. Specifically, the intermediate latent layer representations of the same type of scheduling information terms are extracted within historical scheduling cycles, and the average distance of these latent vectors in the latent space is calculated.

[0050] By analyzing historical stability parameters of the model, including output distribution entropy, maximum a posteriori probability, local prediction gradient norm, and latent space average distance, a model historical stability index is obtained to characterize the predictive stability of the machine learning model. This index takes into account the interrelationships among these parameters. For example, in the process of model historical stability analysis, output distribution entropy, maximum a posteriori probability, local prediction gradient norm, and latent space average distance jointly characterize the stability of the model's predictive behavior from different levels. Among them, output distribution entropy and maximum a posteriori probability reflect the certainty of the model's decision-making from the perspective of the distribution of prediction results, local prediction gradient norm reflects the model's sensitivity to state changes from the perspective of input perturbation response, and latent space average distance characterizes the consistency of the model's representation of the state in continuous scheduling cycles from the perspective of the model's internal representation. When the model output distribution entropy is low and the maximum a posteriori probability is high, it indicates that the model has high consistency in scheduling decisions within the historical cycle. Conversely, a low local prediction gradient norm and a low latent space average distance further indicate that the model has good robustness to input state changes and internal representation stability.

[0051] The sum of the output distribution entropy weight, maximum a posteriori probability weight, local prediction gradient norm weight, and latent space average distance weight is 1. These weights can be obtained through historical data lookup. Specifically, for the same machine learning model within a historical scheduling period, the time series of the four parameters (output distribution entropy, maximum a posteriori probability, local prediction gradient norm, and latent space average distance) are recorded synchronously for each scheduling period. Simultaneously, the model prediction effectiveness evaluation index for that period is obtained. This index is calculated by averaging the scheduling execution success rate and the scheduling instruction rollback rate. Historical stability parameters are standardized to eliminate differences in units and numerical scales. Based on commonly used correlation analysis methods (such as Pearson correlation coefficient), the absolute value of the correlation coefficient between the standardized sequence of each parameter and the prediction effectiveness evaluation value is calculated. This absolute value of the correlation coefficient characterizes the strength of the corresponding parameter's influence on the model's historical stability and prediction reliability. Finally, the absolute values ​​of the correlation coefficients corresponding to the historical stability weights of the model are normalized so that their sum is 1. The normalized results are used as the weights of output distribution entropy, maximum a posteriori probability, local prediction gradient norm, and latent space average distance, respectively.

[0052] Furthermore, the model generalization reliability index is obtained as follows: The model generalization reliability parameters of the machine learning model corresponding to each scheduling information term within the historical scheduling period are obtained. These parameters include historical prediction consistency rate, scheduling suggestion backoff rate, prediction sign inversion density, and distribution deviation response slope. A preset model generalization reliability benchmark set is obtained from the database and compared with the model generalization reliability parameters to obtain the processed values ​​for each model generalization reliability. These processed values ​​are then multiplied by the corresponding weighting factors and superimposed to obtain the model generalization reliability index used to characterize the prediction consistency of the machine learning model across scheduling periods. The model generalization reliability benchmark set includes historical prediction consistency rate benchmark values, scheduling suggestion backoff rate benchmark values, prediction sign inversion density benchmark values, and distribution deviation response slope benchmark values.

[0053] In this embodiment, the historical prediction consistency rate is used to characterize whether the prediction results given by the same machine learning model for the same type of scheduling information term are consistent in adjacent or multiple historical scheduling cycles. It reflects the prediction stability and consistency of the model in the time dimension. It can be achieved by recording the prediction output results of the model for the same scheduling information term in multiple consecutive historical scheduling cycles, discretizing the prediction results into a unified prediction category or prediction direction (e.g., "increase / decrease / no adjustment"), and counting the proportion of the prediction category with the highest frequency in the prediction results as the historical prediction consistency rate. The scheduling suggestion rollback rate is used to characterize the proportion of scheduling suggestions output by the machine learning model that are rejected, withdrawn, or not executed in the actual scheduling process. It reflects the executability and reliability of the model output under real scheduling constraints. It can be achieved by counting the total number of scheduling suggestions output by the model in the historical scheduling cycle, counting the number of scheduling suggestions that were not executed due to conflicts, risk control, or manual / system verification, and the ratio of the total number of scheduling suggestions to the number of scheduling suggestions that were not executed is the scheduling suggestion rollback rate. Prediction sign reversal density is used to characterize whether the prediction direction for the same scheduling objective frequently reverses within consecutive scheduling cycles, reflecting the stability of the model's judgment on state changes. This can be achieved by symbolizing the prediction results within consecutive scheduling cycles (e.g., "+1" or "+2"). The prediction sign reversal density is calculated by dividing the number of reversals by the number of prediction periods (1 or 0). The distribution deviation response slope characterizes the strength of the model output change in response to the degree of deviation when the input data distribution shifts relative to the historical distribution. It reflects the model's sensitivity to distribution changes and its generalization robustness. The deviation of the input distribution from the historical distribution can be measured using statistical distance (such as Kullback-Leibler divergence or mean shift), while simultaneously recording the magnitude of the model output change. A linear fit between the "degree of distribution deviation" and the "magnitude of output change" yields the distribution deviation response slope.

[0054] Historical prediction consistency rate, scheduling suggestion backoff rate, prediction sign inversion density, and distribution deviation response slope are interdependent in evaluating model generalization reliability. Historical prediction consistency rate reflects the model's output consistency over time; a decrease in consistency rate is often accompanied by an increase in prediction sign inversion density, indicating that the model makes unstable judgments about similar states. An increase in prediction sign inversion density further increases the probability of scheduling suggestions being rejected during the coordination and verification phase, thus raising the scheduling suggestion backoff rate. Distribution deviation response slope characterizes the sensitivity of the model's output to changes in the input data distribution. It measures the strength of the model's response to changes in the input distribution; higher generalization reliability leads to a lower historical prediction consistency rate and a higher prediction sign inversion density. These four factors work together to characterize the model's generalization reliability across scheduling cycles from multiple dimensions, including temporal consistency, execution feasibility, directional stability, and robustness.

[0055] A model generalization reliability index for characterizing the prediction consistency of a machine learning model across scheduling cycles is obtained. Specifically, the historical prediction consistency rate (KY) is compared with a preset historical prediction consistency rate benchmark (τKY) to obtain a processed historical prediction consistency rate value (KY / τKY). The scheduling suggestion backoff rate (KT) and prediction sign inversion density (KM) are then compared and normalized with their corresponding scheduling suggestion backoff rate benchmark (τKT) and prediction sign inversion density benchmark (τKM) to obtain processed scheduling suggestion backoff rate values ​​(τKT / KT) and prediction sign inversion density values ​​(τKM / KM). Finally, the distribution deviation response slope (K...) is... V) is compared and normalized with the preset distribution deviation response slope benchmark value (τKV) to obtain the distribution deviation response slope processed value (τKV / KV). Then, it is combined with the historical prediction consistency rate processed value (KY / τKY), the prediction symbol inversion density processed value (τKM / KM), the distribution deviation response slope processed value (τKV / KV), and the scheduling suggestion backoff rate processed value (τKT / KT), and the corresponding credibility weights are introduced and multiplied, and then superimposed to obtain the model generalization credibility index. The credibility weights include the historical prediction consistency rate weight, the scheduling suggestion backoff rate weight, the prediction symbol inversion density weight, and the distribution deviation response slope weight.

[0056] The weights for historical prediction consistency rate, scheduling suggestion rollback rate, prediction sign inversion density, and distribution deviation response slope can be obtained from a database. For example, multiple complete historical scheduling cycle samples can be selected from the database. Each historical scheduling cycle should include at least the model prediction output, the actual scheduling execution result, and the post-scheduling network operation status evaluation result. For each historical scheduling cycle, the actual values ​​of historical prediction consistency rate, scheduling suggestion rollback rate, prediction sign inversion density, and distribution deviation response slope are calculated, and the corresponding scheduling execution effectiveness evaluation index is obtained simultaneously. This scheduling execution effectiveness evaluation index is obtained using a composite evaluation method, that is, simultaneously calculating the scheduling execution success rate and the scheduling instruction rollback occurrence rate. The average value of the deviation of the operational status after scheduling is used as the single evaluation result of scheduling execution effectiveness. Based on correlation analysis methods (such as Pearson correlation coefficient), Pearson correlation coefficients are calculated between the historical prediction consistency rate, scheduling suggestion backoff rate, prediction symbol inversion density, and distribution deviation response slope and the evaluation result of scheduling execution effectiveness. The absolute value of each correlation coefficient is taken to characterize the influence of each parameter on the model's prediction consistency across scheduling cycles. Finally, the absolute values ​​of the correlation coefficients corresponding to the four parameters are normalized so that their sum is 1. The normalized values ​​are used as the weights of historical prediction consistency rate, scheduling suggestion backoff rate, prediction symbol inversion density, and distribution deviation response slope, respectively.

[0057] Furthermore, the execution instructions for each conflict scheduling event are obtained. Specifically, for each scheduling information term in each conflict scheduling information set, the corresponding historical model stability index and generalization reliability index are obtained. These indices are compared with corresponding thresholds, and scheduling information terms above these thresholds are selected as candidate scheduling information terms. The corresponding thresholds include the historical model stability threshold and the generalization reliability threshold. If no candidate scheduling information terms exist in the conflict scheduling information set, none of these terms are executed. If multiple candidate scheduling information terms still exist in the conflict scheduling information set, a comprehensive confidence value is calculated for each term. The scheduling information term with the highest comprehensive confidence value is selected as the conflict scheduling execution instruction for that set. The conflict scheduling execution instructions for each conflict scheduling information set and the scheduling information terms for the normal scheduling information set are then statistically analyzed and executed together to obtain the conflict scheduling execution instructions, which are then output for intelligent scheduling of the public transport network.

[0058] In this embodiment, during the intelligent scheduling of the public transport network, different scheduling information terms may simultaneously apply to the same route, the same time window, or the same transport capacity resource based on different data sources, prediction models, or scheduling objectives, thus forming a set of conflicting scheduling information. If such conflicting scheduling information is not identified and filtered, and is directly executed in parallel, it will lead to mutual cancellation between scheduling instructions, frequent corrections, or fluctuations in transport capacity allocation, reducing the overall scheduling stability and execution controllability.

[0059] In conflict scheduling information sets, a single indicator (such as relying solely on model historical stability or model generalization reliability) is insufficient to comprehensively reflect the actual reliability of scheduling information terms in the current scheduling scenario. By calculating a comprehensive confidence value for candidate scheduling information terms and jointly quantifying and evaluating the model historical stability and model generalization reliability indicators, we can simultaneously characterize the long-term stable performance of the scheduling information term in historical operation, as well as its predictive consistency and adaptability under cross-scheduling cycles and cross-scenario conditions. Thus, when multiple candidate scheduling information terms exist, we can objectively select the scheduling information term with the highest overall reliability and lowest execution risk, effectively reducing the probability of misscheduling and improving the robustness of scheduling decisions in conflict scenarios.

[0060] Further, the comprehensive confidence value is calculated using the following method: The model's historical stability index and historical stability threshold are compared to obtain the model's historical stability margin. The model's generalization confidence index and generalization confidence threshold are then compared to obtain the model's generalization confidence margin. The model's historical stability margin and generalization confidence margin are combined and compared with a preset benchmark margin set to obtain various margin processing values. Each margin processing value includes the model's historical stability margin processing value (obtained by dividing the model's historical stability margin by the benchmark value) and the model's generalization confidence margin processing value (obtained by dividing the model's generalization confidence margin by the benchmark value). These margin processing values ​​are then multiplied by their corresponding margin weight sets and superimposed to obtain the comprehensive confidence value used to characterize the execution reliability of the scheduling information term. The margin weight set includes the model's historical stability benchmark margin weight and the model's generalization confidence benchmark margin weight.

[0061] In this embodiment, the machine learning model is used to generate scheduling information terms based on the bus network operation status data. Its model structure includes an input layer, a feature encoding layer, and an output layer connected in sequence. The input layer receives state feature parameters corresponding to the current bus network operation status, including at least the route passenger load field, station congestion field, vehicle arrival time deviation field, vehicle turnaround status field, and departure interval control field. The feature encoding layer performs feature mapping processing on the multi-dimensional operation status features to establish the correspondence between the network operation status and the scheduling strategy. The output layer outputs scheduling information terms matching the operation status. These scheduling information terms include at least one or more of the following: additional departures, delayed departures, section skipping, detours, or capacity adjustments. The machine learning model is at least one of the following: MLP (Multilayer Perceptron), DNN (Deep Neural Network), XGBoost (Extreme Gradient Boosting Tree Model), or LightGBM (Lightweight Gradient Boosting Machine Model).

[0062] The training process of the machine learning model includes: acquiring snapshot data of the bus network operation status within historical scheduling cycles and the corresponding actual scheduling execution results to construct a training sample set; using the state feature fields in the network operation status snapshots as model input features and the actual scheduling strategies adopted within historical scheduling cycles as supervision labels, updating the model parameters through iterative training to enable the model to learn the decision distribution relationship of generating scheduling information terms under different network operation states. During training, the magnitude of changes in the model output distribution, prediction consistency rate, parameter update magnitude, and prediction confidence interval width are recorded as parameters of model historical stability; simultaneously, the output consistency rate, strategy reversal frequency, and prediction distribution drift degree of the model under different time periods and different route combinations are recorded as parameters of model generalization reliability.

[0063] In practical applications, when a conflict scheduling scenario is detected, the current bus network operation status characteristics are input into the trained machine learning model to generate a corresponding set of scheduling information terms. The model's historical stability index is calculated based on the model's historical stability parameters, and the model's generalization reliability index is calculated based on the model's generalization reliability parameters. The model's historical stability index and the model's generalization reliability index are compared with preset thresholds to filter the scheduling information terms generated by the machine learning model. Scheduling information terms generated by models with insufficient stability or insufficient generalization reliability are removed, thereby obtaining conflict scheduling execution instructions for execution.

[0064] In this embodiment, as Figure 6 As shown, Figure 6This diagram illustrates the stability and reliability distribution of candidate scheduling information terms in the intelligent scheduling method for public transport networks based on machine learning models provided in this application. The horizontal axis represents the historical stability index of the model, which can be obtained by comparing and weighting stability parameters such as output distribution entropy, maximum a posteriori probability, local prediction gradient norm, and latent space average distance with the benchmark set. The vertical axis represents the generalization reliability index of the model, which can be obtained by comparing and weighting generalization parameters such as historical prediction consistency rate, scheduling suggestion backoff rate, prediction sign inversion density, and distribution deviation response slope with the benchmark set. Each scatter point Ti represents the (stability, generalization) score position of a scheduling information term within the conflict set in the current evaluation period. The vertical / horizontal dashed lines in the figure represent the stability threshold and the generalization reliability threshold, respectively, dividing the plane into four regions: the upper right region represents high-reliability candidates that are "stable and have high generalization reliability," the lower left region represents low-reliability terms that are "insufficient in both stability and generalization" and are usually excluded, and the upper left / lower right regions correspond to risky candidates that are "good in generalization but have large short-term fluctuations" or "stable in output but insufficient in adapting to new scenarios," respectively, requiring more careful inclusion or rollback control. The star (★) in the figure represents the highest-scoring term selected from the candidate terms that have passed the threshold, which is the conflict scheduling execution instruction that is finally determined and issued for the conflict scheduling information set.

[0065] The historical stability margin weights and generalization reliability margin weights of the model can be obtained from a database. Specifically, the following method is used: A historical scheduling sample set with completed execution and complete result feedback is selected from the database. For each scheduling information term, its historical stability margin and generalization reliability margin are calculated. Corresponding objective execution performance evaluation indicators are extracted, such as the frequency with which the scheduling execution success rate is corrected in subsequent cycles. Statistical correlation analysis is used to calculate the absolute value of the correlation coefficient between the historical stability margin and the execution performance evaluation indicators, yielding the historical stability margin correlation coefficient. Similarly, the absolute value of the correlation coefficient between the generalization reliability margin and the execution performance evaluation indicators is calculated, yielding the generalization reliability margin correlation coefficient. The absolute values ​​of these two correlation coefficients are used as the contribution strength representation values ​​of the historical stability margin and the generalization reliability margin to the scheduling execution reliability, respectively. Both are then normalized so that the sum of the normalized historical stability margin weights and generalization reliability margin weights is 1.

[0066] Furthermore, the merging and exclusion of scheduling information terms in subsequent scheduling cycles are controlled. Specifically, each conflicting scheduling execution instruction is managed with a scheduling effectiveness identifier. Executed conflicting scheduling execution instructions are marked as effective scheduling terms, while unexecuted conflicting scheduling execution instructions are marked as ineffective scheduling terms. After generating a network operation status snapshot and entering the corresponding scheduling cycle, scheduling coordination freeze constraints are established to control the merging and exclusion of scheduling information terms in subsequent scheduling cycles. The scheduling coordination freeze constraints include: each scheduling information term must be generated based on the same network operation status snapshot, and terms already marked as effective scheduling terms do not participate in duplicate / overlap detection.

[0067] In this embodiment, as Figure 7 As shown, Figure 7 The flowchart for screening and determining conflict scheduling execution instructions in the intelligent bus network scheduling method based on machine learning models provided in this application embodiment is as follows: For each scheduling information term in the set, the corresponding historical model stability index and model generalization reliability index are obtained, and these indices are compared with their respective thresholds for screening. Only scheduling information terms that are simultaneously higher than both the historical model stability threshold and the model generalization reliability threshold are included in the candidate scheduling information term set. When there are no candidate scheduling information terms in the conflict scheduling information set, the system directly determines that no scheduling information terms will be executed for that conflict scheduling information set. When the number of candidate scheduling information terms is greater than one, the system further calculates a comprehensive confidence value for each candidate scheduling information term and compares them according to the magnitude of the comprehensive confidence value, selecting the scheduling information term with the highest comprehensive confidence value as the conflict scheduling execution instruction for that conflict scheduling information set. All conflict scheduling execution instructions are output and executed together with the scheduling information terms in the normal scheduling information set for intelligent bus network scheduling.

[0068] "Incorporation" refers to introducing scheduling information terms that meet the scheduling coordination freeze constraints and originate from a new round of scheduling calculations or external real-time operational feedback into the candidate scheduling information set of subsequent scheduling cycles for analysis and decision-making. "Exclusion," on the other hand, refers to removing scheduling information terms that do not meet the incorporation conditions, preventing them from participating in the scheduling generation process of subsequent scheduling cycles. By introducing scheduling effectiveness flag management and scheduling coordination freeze constraints in subsequent scheduling cycles, the orderly evolution of scheduling information terms in both time and state dimensions can be effectively achieved. This avoids executed scheduling instructions being repeatedly evaluated, conflicting, or overwritten in subsequent cycles. Furthermore, controlling the incorporation and exclusion of scheduling information terms based on a unified network operation status snapshot helps ensure the temporal consistency and state interpretability of scheduling decisions. This prevents historical scheduling results from implicitly interfering with subsequent scheduling judgments and provides a stable and continuous decision-making environment for machine learning models running across scheduling cycles.

[0069] The second embodiment, based on the first embodiment, obtains each conflict scheduling execution instruction and can also be implemented in the following way: It acquires the data upload stability parameters of the data upload nodes corresponding to each scheduling information term. These data upload stability parameters include transmission interval variance, uplink timing jitter intensity, and short-term uplink interruption density, and analyzes to obtain data upload stability indices. Based on the data upload stability indices and combined with the model's historical stability indices and model generalization reliability indices, if a scheduling information term in the conflict scheduling information set has a data upload stability index above a preset data upload stability threshold, a model's historical stability index above a preset data upload stability threshold, and a model's generalization reliability index above a preset model generalization reliability threshold, then that scheduling information term is considered a candidate scheduling information term. Where, when the conflict scheduling information set contains a scheduling information term, the data upload stability index of that scheduling information term is above a preset data upload stability threshold, the model's historical stability index is above a preset data upload stability threshold, and the model's generalization reliability index is above a preset model generalization reliability threshold, then that scheduling information term is considered a candidate scheduling information term. If no candidate scheduling information terms exist in the conflict scheduling information set, no scheduling information terms within that conflict scheduling information set will be executed. If multiple candidate scheduling information terms still exist in the conflict scheduling information set, a comprehensive confidence value is calculated for each candidate scheduling information term. The comprehensive confidence value is then averaged with the data upload stability index to obtain a comprehensive data upload index. This is used to iterate through each candidate scheduling information term to obtain the comprehensive data upload index for each candidate scheduling information term. The candidate scheduling information term corresponding to the maximum comprehensive data upload index is selected as the conflict scheduling execution instruction for that conflict scheduling information set. The conflict scheduling execution instructions corresponding to each conflict scheduling information set and the scheduling information terms of the normal scheduling information set are then statistically analyzed and executed together to obtain each conflict scheduling execution instruction, which is then output for intelligent scheduling of the public transport network.

[0070] Transmission interval variance characterizes the dispersion of the time interval between two adjacent data uplink uploads from the same data upload node within a continuous scheduling period. Specifically, it can be obtained by recording the timestamps of multiple data reports from the upload node within a unit scheduling period, calculating the difference between adjacent timestamps, and using the time series variance calculation method in the existing technology. Uplink timing jitter intensity characterizes the fluctuation of the actual arrival time order of data packets at the scheduling center relative to the theoretical transmission order. It can be obtained by comparing the relative offset between the data packet transmission timestamp and the reception timestamp, and by statistically analyzing the difference in continuous upload delay based on the delay jitter calculation method commonly used in existing communication networks. Uplink link short-term interruption density characterizes the frequency of short-term unreachability or continuous data packet loss events in the data upload link within a preset observation time window. It can be obtained by statistically analyzing the number of events of link disconnection, retransmission failure, or continuous failure to receive acknowledgment feedback by the upload node within a unit time, and then performing density processing in combination with the window duration.

[0071] In data upload stability analysis, transmission interval variance reflects the temporal uniformity of upload rhythm, uplink timing jitter intensity reflects the transient fluctuation characteristics of data arrival order and delay, and short-term interruption density of upload links reflects the structural discontinuity risk of link availability. These three factors exhibit a synergistic influence in actual operation: when a short-term interruption occurs in the upload link, it often directly amplifies the transmission interval variance, causing abrupt changes in adjacent upload time intervals. It also triggers packet buffering, reordering, or concentrated retransmission, thereby significantly increasing the uplink timing jitter intensity. Conversely, even if the link is not completely interrupted, if network congestion or unstable node scheduling leads to a continuous increase in uplink timing jitter, it will indirectly increase the transmission interval variance. Therefore, data upload stability indicators are not determined by a single parameter, but rather by joint analysis and weighted synthesis of the three types of parameters to characterize the overall stability level of the upload link in terms of temporal continuity, sequential consistency, and reachability reliability.

[0072] The data upload stability index is obtained by comparing the transmission interval variance (WS), uplink timing jitter intensity (WQ), and upload link short-term interruption density (WM) with preset baseline values ​​for transmission interval variance (τWS), uplink timing jitter intensity (τWQ), and upload link short-term interruption density (τWM) to obtain the data upload stability processing values. Each data upload stability processing value includes the transmission interval variance processing value (τWS / WS), the uplink timing jitter intensity processing value (τWQ / WQ), and the upload link short-term interruption density processing value (τWM / WM). Each data upload stability processing value is multiplied by its corresponding data upload processing weight and then superimposed to obtain the data upload stability index. The data upload processing weights include the transmission interval variance weight (γ1), the uplink timing jitter intensity weight (γ2), and the upload link short-term interruption density weight (γ3).

[0073] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent scheduling of public transport networks based on machine learning models, characterized in that, Includes the following steps: Receive network sensing data, and when an abnormal evolution trend is detected in the network sensing data, generate a snapshot of the network operation status and send it to the scheduling information coordination and processing unit to generate various scheduling information entries; The overlapping detection is performed on each scheduling information term, the overlapping relationship matrix is ​​output, and the normal scheduling information set and each conflict scheduling information set are obtained. The conflict scheduling information set includes each conflict scheduling information term with a certain overlapping relationship. The training parameters of the machine learning model for each conflict scheduling information term are obtained. The machine learning model is used to generate scheduling information terms based on the bus network operation status data. The historical stability index and generalization reliability index of the machine learning model are analyzed and obtained. Based on these, the execution instructions for each conflict scheduling are selected to execute intelligent scheduling of the bus network. The training parameters of the machine learning model include the historical stability parameters and generalization reliability parameters of the machine learning model. The historical stability parameters include output distribution entropy, maximum posterior probability, local prediction gradient norm, and latent space average distance. The generalization reliability parameters include historical prediction consistency rate, scheduling suggestion backoff rate, prediction sign inversion density, and distribution deviation response slope. The system manages the scheduling effectiveness identifiers of all executed conflict scheduling instructions, and controls the incorporation and exclusion of scheduling information entries in subsequent scheduling cycles based on these identifiers to achieve intelligent scheduling of the public transport network.

2. The intelligent bus network scheduling method based on machine learning model as described in claim 1, characterized in that: The specific detection method for the aforementioned abnormal evolution trend is as follows: Within a preset scheduling period, the state field evolution analysis is performed on the time-aligned network sensing data to evaluate the magnitude and direction of change of each network state field between adjacent scheduling periods, and to obtain the magnitude and direction of change results of each network state field. When the change magnitude of at least one network status field exceeds the corresponding preset stable range and the change direction is determined to be positive evolution or oscillating evolution, and / or when the network status fields of multiple lines show coordinated abnormal offset within the same scheduling cycle and the change direction is determined to be positive evolution or oscillating evolution, it is determined that the network sensing data has an abnormal evolution trend.

3. The intelligent bus network scheduling method based on machine learning model as described in claim 1, characterized in that: The specific method for generating the snapshot of the network operation status is as follows: Within a preset scheduling period, receive network sensing data from different functional modules, and map the network sensing data to a unified scheduling time window according to timestamps; The mapped wire network sensing data is time-aligned to remove historical and forward data that do not belong to the current scheduling cycle. The time-aligned multi-source network sensing data is subjected to state abstraction processing to extract unified network status fields, which include departure interval field, section operation density field, station load change field and capacity configuration status field. Based on the unified network status field, a basic dataset of network operation status is constructed, and this basic dataset of network operation status is solidified into a snapshot of the network operation status for the corresponding scheduling period.

4. The intelligent bus network scheduling method based on machine learning model as described in claim 1, characterized in that: The specific method for outputting the overlap relationship matrix is ​​as follows: Each scheduling information term is parsed using the resource identifier dimension, which includes the target dimension, the route dimension, and the time window dimension. Based on the dimension of the target object, construct a mapping table of the interaction relationships between scheduling information terms; When any two scheduling information terms intersect in at least one dimension of the target object, they are marked as overlapping; The overlap relationships between all scheduling information terms are represented in matrix form to obtain the overlap relationship matrix.

5. The intelligent bus network scheduling method based on a machine learning model as described in claim 1, characterized in that: The specific method for obtaining the normal scheduling information set and the conflict scheduling information sets is as follows: Based on the aforementioned overlap matrix, association analysis is performed on the scheduling information terms, and scheduling information terms that do not overlap with other scheduling information terms are assigned to the normal scheduling information set. The scheduling information entries that have at least one overlapping relationship with each other are grouped into the same conflict scheduling information set, thereby obtaining each conflict scheduling information set; Each conflict scheduling information set contains scheduling information entries that share a common conflict target.

6. The intelligent bus network scheduling method based on a machine learning model as described in claim 1, characterized in that: The specific method for obtaining the historical stability index of the machine learning model is as follows: Obtain the historical stability parameters of the machine learning model corresponding to each scheduling information term within the historical scheduling period; Obtain a preset model historical stability benchmark set and compare it with the model historical stability parameters to obtain the historical stability processing value of each model. Based on the historical stability processing value of each model, introduce the corresponding weights, multiply them, and then superimpose them to obtain the model historical stability index used to characterize the predictive stability of the machine learning model. The historical stability benchmark set of the model includes the output distribution entropy benchmark value, the maximum a posteriori probability benchmark value, the local prediction gradient norm benchmark value, and the latent space average distance benchmark value.

7. The intelligent bus network scheduling method based on a machine learning model as described in claim 1, characterized in that: The method for obtaining the model generalization reliability index is as follows: Obtain the model generalization reliability parameters of the machine learning model corresponding to each scheduling information term within the historical scheduling period; Obtain the preset model generalization confidence benchmark set in the database and compare it with the model generalization confidence parameters to obtain the generalization confidence processing value of each model. Introduce the generalization confidence processing value of each model into the corresponding weighting factor, multiply it, and then superimpose it to obtain the model generalization confidence index used to characterize the prediction consistency of the machine learning model across scheduling cycles. The model generalization reliability benchmark set includes historical prediction consistency rate benchmark, scheduling suggestion backoff rate benchmark, prediction symbol inversion density benchmark, and distribution deviation response slope benchmark.

8. The intelligent scheduling method for public transport networks based on machine learning models as described in claim 1, characterized in that: The specific method for obtaining the conflict scheduling execution instructions is as follows: For each scheduling information term in each conflict scheduling information set, obtain the corresponding model historical stability index and model generalization reliability index; The historical stability index and the generalization reliability index of the model are compared with the corresponding thresholds, and scheduling information terms that are above the corresponding thresholds are selected as candidate scheduling information terms. The corresponding thresholds include the historical stability threshold and the generalization reliability threshold. If there are no candidate scheduling information terms in the conflict scheduling information set, then no scheduling information terms in that conflict scheduling information set will be executed. When there are still multiple candidate scheduling information terms in the conflict scheduling information set, calculate the comprehensive confidence value for each candidate scheduling information term. The scheduling information term with the highest comprehensive confidence value is selected as the conflict scheduling execution instruction for the conflict scheduling information set. The conflict scheduling execution instructions corresponding to each conflict scheduling information set and the scheduling information terms of the normal scheduling information set are then counted and used as the conflict scheduling execution instructions, which are then output for intelligent scheduling of the public transport network.

9. The intelligent scheduling method for public transport networks based on machine learning models as described in claim 8, characterized in that: The specific method for calculating the comprehensive confidence value is as follows: The model historical stability margin is obtained by difference processing between the model historical stability index and the model historical stability threshold. The model generalization confidence margin is obtained by difference processing between the model generalization confidence index and the model generalization confidence threshold. The model historical stability margin and the model generalization confidence margin are combined and compared with the preset benchmark margin set to obtain each margin processing value. Each margin processing value is introduced into the corresponding margin weight set, multiplied, and then superimposed to obtain the comprehensive confidence value used to characterize the execution reliability of the scheduling information term. The baseline margin set includes the historical stability margin baseline value and the generalization confidence margin baseline value of the model; The margin weight set includes the historical stability baseline margin weight and the generalization reliability baseline margin weight.

10. The intelligent bus network scheduling method based on a machine learning model as described in claim 1, characterized in that: The specific method for controlling the inclusion and exclusion of scheduling information terms in subsequent scheduling cycles is as follows: Each conflict scheduling execution instruction is managed by scheduling effectiveness identifiers. Each executed conflict scheduling execution instruction is marked as a scheduling effective entry, and each unexecuted conflict scheduling execution instruction is marked as a scheduling ineffective entry. After generating a snapshot of the network operation status and entering the corresponding scheduling cycle, a scheduling coordination freeze constraint is established to control the incorporation and exclusion of scheduling information entries in subsequent scheduling cycles. The scheduling coordination freeze constraints include: each scheduling information term must be generated based on the same network operation status snapshot, and terms that have been marked as effective for scheduling will not participate in duplicate or overlapping detection.

Images